Best ratio CPU usage / level

[Peperoni]

Hello,

We are thinking about doing a chess engine in C++ 20 that would run on mobile devices.
CPU is then a concern and we are trying to find the best approch to have a good ratio CPU usage/ level of the engine.
Do you guys think something with neural network weights would be lighter than the classical intensive CPU approch?

Thank you

[smatovic]

[...]
Do you guys think something with neural network weights would be lighter than the classical intensive CPU approch?
[...]

I doubt that, NNUE neural networks are currently implemented via vector-units (SSE, AVX2, NEON), which are more efficient than the CPU ALUs but I doubt that you will safe power usage with NNUE compared to classic HCE in an NPS or even Elo metric. SoCs have usual an dedicated neural engine, a TPU or MMAC unit, and a programmable GPU, might be worth to check out with an Lc0 CNN kind of approach.

--
Srdja